Short-term Traffic Flow Prediction Based On The Long Short Term Memory Network Model

Urbanization is an inevitable trend of social development.With the develop-ment of urbanization,urban traffic will produce traffic congestion,frequent traffic accidents,air pollution,waste of oil resources,noise pollution and other issues.In order to improve the traffic situation,it is not enough to only rely on improving the infrastructure construction,but also need scientific traffic planning and man-agement,which is the general trend of the development of urban public transport in the future.Accurate and real-time traffic flow prediction is an important part of it.It can predict the traffic flow situation in the future period.For traffic managers,it can be divided and regulated in advance to alleviate urban traf-fic congestion.For individual travelers,it can improve efficiency and save time and economic cost.For society,it can improve the utilization rate of economy and natural resources and promote the harmonious development of society.In addition,with the development of traffic information collection equipment and computing and transmission technology,traffic data can be acquired in real time and diversified,which also lays a foundation for model training.Thus it becomes a research hotspot to find a suitable method to predict the traffic flow changes in a short time.Due to the obvious periodicity of traffic flow data,this paper selects the long short term memory(LSTM)network model as the basic model for traffic flow prediction,but the prediction results of the single model are not good.After that,this paper improves the model and proposes the classified LSTM model to predict the traffic flow time series.Moreover,considering the influence of spatial factors on traffic flow,this paper uses convolution LSTM(ConvLSTM)model to predict the time-space sequence of trafic flow,which can capture the spatial characteristics of traffic flow data,take into account the advantage of memory in LSTM model and reduce the complexity of convolution neural net-work.Meanwhile,this paper combines the genetic algorithm with ConvLSTM model.Genetic algorithm is used to filter the spatiotemporal nodes to reduce the dimension first.Then ConvLSTM model is used to predict the model,so as to improve the prediction efficiency.The classified LSTM model combines the clustering algorithm with the LSTM model.The model classifies the traffic flow data properly,and then establishes the corresponding LSTM submodel according to different types of data.First of all,this paper classifies the data into two categories:violent and gentle by K-means clustering.Then,for these two types of data samples,the results of the parameters in the main LSTM model are taken as the initial parameters of the submodel,and the corresponding LSTM submodels are trained respectively.Finally,the two submodels are fused to predict the test set.The ConvLSTM model introduces convolution calculation to extract the spa-tial characteristics of traffic flow for prediction.Firstly,the space range is deter-mined.Then,in order to reduce the data dimension,genetic algorithm is used to filter the spatial nodes.Taking the traffic flow of each spatiotemporal node as its image value,we use convolution to extract spatial features.On the basis of fully connected LSTM model,the calculation method of input to state and state to state is changed to convolution calculation.In the aspect of data experiment,this paper focuses on the traffic flow of VDS 312098 collected by PEMs system in Sacramento County,California in USA.And we predict the traffic flow of the point after 5 minutes.MAE,RMSE and MRE are used as evaluation criteria to compare the prediction results of ARIMA,RNN,LSTM and its two extended models.The final results show that the three error indexes of the two extended models are all reduced.